SaGE: Evaluating Moral Consistency in Large Language Models
International Institute of Information Technology Hyderabad (IIITH)(国际信息技术学院海得拉巴分校(IIITH));University of Maryland Baltimore County (UMBC)(马里兰大学巴尔的摩县分校(UMBC))
尽管近期的研究进展展现了大语言模型(LLMs)在对话系统中令人瞩目的能力,但我们发现,即便是最先进的LLMs,其生成内容也存在道德不一致的问题,这对它们的可靠性(以及整体可信度)提出了质疑。此前的LLM评估研究主要侧重于构建基准真值数据,以衡量模型在特定任务上的准确率。但对于往往缺乏公认答案的道德场景而言,模型回答的一致性对其可靠性至关重要。为解决这一问题,我们提出了一种名为语义图熵(Semantic Graph Entropy, SaGE)的信息论度量方法,该方法以“经验法则”(Rules of Thumb, RoTs)概念为基础,用于衡量模型的道德一致性。RoTs是模型习得的抽象原则,可有效解释其决策策略。为此,我们构建了道德一致性语料库(Moral Consistency Corpus, MCC),包含5万个道德问题、LLM对这些问题的回答,以及这些模型遵循的RoTs。此外,为说明SaGE的泛化性,我们用它研究了LLM在两个流行数据集——TruthfulQA和HellaSwag上的一致性。我们的结果表明,任务准确率和一致性是相互独立的问题,亟需对这些问题开展进一步研究。
英文摘要:
Despite recent advancements showcasing the impressive capabilities of Large Language Models (LLMs) in conversational systems, we show that even state-of-the-art LLMs are morally inconsistent in their generations, questioning their reliability (and trustworthiness in general). Prior works in LLM evaluation focus on developing ground-truth data to measure accuracy on specific tasks. However, for moral scenarios that often lack universally agreed-upon answers, consistency in model responses becomes crucial for their reliability. To address this issue, we propose an information-theoretic measure called Semantic Graph Entropy (SaGE), grounded in the concept of "Rules of Thumb" (RoTs) to measure a model's moral consistency. RoTs are abstract principles learned by a model and can help explain their decision-making strategies effectively. To this extent, we construct the Moral Consistency Corpus (MCC), containing 50K moral questions, responses to them by LLMs, and the RoTs that these models followed. Furthermore, to illustrate the generalizability of SaGE, we use it to investigate LLM consistency on two popular datasets -- TruthfulQA and HellaSwag. Our results reveal that task-accuracy and consistency are independent problems, and there is a dire need to investigate these issues further.
The Moral Consistency Pipeline: Continuous Ethical Evaluation for Large Language Models
SWAT Laboratory, Polytechnique Montréal(蒙特利尔大学SWAT实验室);Concordia Institute for Information Systems Engineering, Concordia University(Concordia大学信息系统工程研究所)
Explaining Intrinsic Moral Self-Correction with Mechanistic Interpretability
Graduate Institute of Communication Engineering, National Taiwan University, Taipei, Taiwan(通讯工程研究院,国立台湾大学);MediaTek Research, Taipei, Taiwan(联发科研究,台北,台湾);Department of Electrical Engineering, National Taiwan University, Taipei, Taiwan(电气工程系,国立台湾大学);Department of Electrical Engineering, National Tsing Hua University, Hsinchu, Taiwan(电气工程系,国立清华大学);AI Research Center (AINTU), National Taiwan University, Taipei, Taiwan(人工智能研究中心(AINTU),国立台湾大学)
Discerning What Matters: A Multi-Dimensional Assessment of Moral Competence in LLMs
Machine Intelligence and Normative Theory Lab, Australian National University, Acton, ACT 2601, Australia(澳大利亚国立大学机器智能与规范理论实验室);Digital Media Research Centre GenAI Lab, Queensland University of Technology, Kelvin Grove, QLD 4012, Australia(昆士兰科技大学数字媒体研究中心GenAI实验室)
School of Philosophy (Political Philosophy) Renmin University of China(哲学学院(政治哲学)中国人民大学);School of Government and Policy Johns Hopkins University(政府与政策学院约翰霍普金斯大学)
MFTCXplain: A Multilingual Benchmark Dataset for Evaluating the Moral Reasoning of LLMs through Multi-hop Hate Speech Explanation
University of Southern California(南加州大学);São Paulo State University(圣保罗州立大学);Saarland University(萨尔兰大学);University of Melbourne(墨尔本大学);Howard University(霍华德大学);Portland State University(波特兰州立大学);Leiden University(莱顿大学)
CoMMET: A Psychologically Grounded Benchmark for Evaluating Theory of Mind in Multimodal LLMs
Agency for Science, Technology and Research (A*STAR)(科技研究局);Nanyang Technological University(南洋理工大学);Hong Kong University of Science and Technology (Guangzhou), China(香港科技大学(广州),中国)
MoReBench: Evaluating Procedural and Pluralistic Moral Reasoning in Language Models, More than Outcomes
University of Washington(华盛顿大学);New York University(纽约大学);Scale AI;Harvard University(哈佛大学);University of Michigan(密歇根大学);UNC Chapel Hill(北卡罗来纳大学教堂山分校);Center for AI Safety(人工智能安全中心);Stanford University(斯坦福大学);MIT(麻省理工学院);University of Oxford(牛津大学)
Antisocial Analagous Behavior, Alignment and Human Impact of Google AI Systems: Evaluating through the lens of modified Antisocial Behavior Criteria by Human Interaction, Independent LLM Analysis, and AI Self-Reflection
Probing the contents of semantic representations from text, behavior, and brain data using the psychNorms metabase
Faculty of Psychology, University of Basel(巴塞尔大学心理学系);Center for Adaptive Rationality, Max Planck Institute for Human Development(马克斯·普朗克人类发展研究所适应性理性中心);School of Psychology, UNSW(新南威尔士大学心理学学院)